Evidence

Top-5 by training compute, language models only — narrower than Epoch's site-wide 'frontier' definition (top 10 at time of release). Epoch reports the rate is stable across N=5/10/15 with no statistically significant difference. Headline: 5x/year since 2020 (90% CI 4-6x), doubling every 5.2 months, 0.7 OOM/year; the top-5 trend has grown ~10,000x since 2020. IMPORTANT FRESHNESS CAVEAT: the data bundle behind Epoch's chart ends at 2025-08-06 even though the page banner reads Feb 2026, so the newest point here is ~11 months old — a publication lag, not the censoring problem that afflicts the record-run series. Epoch's own view is that this pace holds ~1-2 years from Sept 2025 before training-duration ceilings (~9 months) and datacenter lead times bite. Methodology detail sits in a private Colab and could not be verified.

The 8-month figure is pretraining-only and backward-looking (2012-2023 fit). Epoch's own decomposition: 60-95% of capability gains came from compute and data, algorithms only 5-40%. A Nov 2025 MIT preprint attributes ~91% of measured efficiency gains to two one-off scale-dependent transitions (LSTM-to-Transformer, Kaplan-to-Chinchilla), cautioning against treating the rate as a steady state. The post-training era complicates measurement further: Epoch (Aug 2025) estimates reasoning-model RL delivered a ~10x compute-equivalent gain (range 1x-100x) in verifiable domains — not captured by the pretraining halving rate. Epoch's live trends dashboard shows ~7.6-month halving, but its provenance/update cadence is undocumented; the Mar 2024 paper remains the citable number.

Task-suite methodology changed v1.0->v1.1 (Jan 2026); cross-version comparisons need care. Measurements above ~16h are near the edge of what the current suite resolves. Single research org, not independently audited. Eval-gaming is becoming the dominant uncertainty: METR's GPT-5.6 Sol eval (Jun 2026) measured 11.3h under standard scoring but 71h-270h+ depending on how its record 55% detected-cheating rate is handled. METR added a companion 'expenditure horizon' metric (cost-based, Jul 2026) — not yet a leaderboard series.

This is the hardest Tier 1 variable to measure, but every takeoff model turns on it. Evidence hierarchy applies hard here: lab-leader statements are intention data and self-reported speedups receive a confidence discount. METR's early-2025 randomized study found a 19% slowdown in adjacent open-source work; its later experiment was inconclusive because of selection and measurement limits. The field still has no clean causal estimate of frontier AI-R&D acceleration.

Total capex is NOT AI-specific — no company breaks it out; third-party estimates of the AI share range 40-75% (JPMorgan ~70% for 2025; Dell'Oro: accelerator silicon alone is ~1/3) depending on definition. Series here is cash 'purchases of property and equipment' EXCLUDING finance leases — earnings-call figures often include them (this basis gap largely explains ~$411B vs Epoch's ~$500B for 2025). Oracle adds ~$248B of off-balance-sheet lease commitments and reports on a May-end fiscal year; Microsoft on June-end — quarterly alignment is by calendar quarter of period end. History values are aggregator-compiled (stockanalysis.com), spot-checked against press figures, not line-by-line 10-Q reconciled.

Three power measures get conflated constantly: IT power (compute equipment draw — used here), facility power (20-50% higher, cooling/overhead), and grid-connection capacity. Operator claims run ~2x above third-party verification (Musk claimed Colossus 2 at '2 GW' in Jan 2026; satellite analysis then showed ~350 MW cooling; Epoch measures 946 MW in Jul 2026). Announced figures: sum of top-5 campus targets ~20-25 GW (Stargate >9 GW by 2029, Meta Hyperion 5 GW, Rainier ~2.2-4.6 GW (sources conflict), Colossus 2 target 2 GW, Prometheus >=1 GW). GPT-4-era baseline point is an unverified secondary estimate. Training-vs-inference cluster labels are operator-reported snapshots. ANNOUNCED SERIES: operator claims plotted at the date each was made, with a bar running to the delivery date promised for it; an X marks an announcement that was cancelled. Operators almost never say whether a figure is IT or facility power, so every announced figure here is UNSPECIFIED on that axis and is not strictly like-for-like with the measured IT-power line. Sources are Musk's X posts, Zuckerberg's Threads posts, and company newsroom releases; xAI has never published a datacenter power figure on its own site. Deliberately EXCLUDED: Stargate's ~10 GW is a multi-site PROGRAM total, not a campus. Project Rainier's widely-cited 2.2 GW traces to Indiana state officials and site plans — neither AWS nor Anthropic has ever attached a GW figure to that campus. Microsoft has never published a power figure for either Fairwater site; the circulating 3.3 GW is Epoch's own independent estimate, which Epoch itself calls 'somewhat more speculative than our other estimates'. That absence is itself informative: two of the four largest builders disclose no campus power at all. Note the asymmetry in the announced series: every revision is upward (Hyperion 2 GW to 5 GW) except one — Stargate Abilene, capped back at 1.2 GW in Mar 2026 when a ~600 MW expansion was scrapped after grid-interconnection delays. That is the physical constraint this stage exists to watch, showing up in an operator's own plans rather than in an analyst's forecast.